Are Data Centres the Bottleneck for Businesses Scaling AI?

A contradiction is emerging in the technology sector. While investment pours into building the most powerful data centre infrastructure in history, most companies cannot seem to leverage it. The advent of AI has created a paradox where the hardware is abundant but the application is scarce.

According to research from Goldman Sachs, global data centres are predicted to consume 165% more electricity by 2030. In contrast, a study from Accenture finds that only 8% of enterprises have successfully scaled AI projects beyond the pilot phase. This highlights a growing gap between infrastructure capability and business execution.

Data centre power consumption

Current data centres consume approximately 55GW of electricity globally. Business applications like email accounts for about a third of that power while AI workloads represent just 14%. Goldman Sachs projects a notable change by 2027 with total consumption rising to 84GW and AI accounting for over a quarter of all power usage.

The power disparity between AI and traditional computing is significant. A single ChatGPT query, for instance, uses almost 10 times the electricity of a Google search at 2.9 watt-hours. AI server racks consume considerably more electricity than their traditional cloud counterparts due to the intensive demands of machine learning models.

For Europe, this presents a particular challenge.

“Inflecting power demand is monumentally important, because it’s been declining for 15 years in Europe,” says Alberto Gandolfi, Managing Director, Equity Research at Goldman Sachs. 

The continent is estimated to require a data centre pipeline of around 170GW, which is equivalent to one-third of its current total electricity consumption.

To meet this demand, substantial investment is needed. According to Goldman Sachs, US utilities alone must spend US$50bn on new generation capacity solely for data centres, while worldwide grid upgrades might cost US$720bn by 2030.

Frank Long, a Vice President at the Goldman Sachs Global Institute suggests: “Retrofitting existing facilities to support these massive jumps in power density is becoming complex and compromised. We will need new, purpose-built AI infrastructure to power the next generation.”

The enterprise scaling crisis

While the infrastructure race continues, most companies are struggling with AI deployment. Accenture's survey of 2,000 executives from billion-dollar companies reveals a disconnect between ambition and reality. The report categorises businesses into three groups:

Within the top tier, only 8% are true front-runners that have managed to scale multiple AI initiatives across their operations. The primary obstacles are not just technical but also human. Poor data readiness, particularly with unstructured data, is a major hurdle. Legacy IT systems create additional friction and building multi-disciplinary teams proves difficult.

“We are writing the playbook for how to be the most AI-enabled, client-focused professional services company in the world,” says Julie Sweet, Chair and CEO at Accenture.

A structured approach to AI foundation

Research from KPMG indicates that overcoming these scaling challenges requires a structured approach. Its analysis identifies five critical strategies for businesses to succeed with AI.

The first is to adopt a phased approach focusing on high-impact, low-risk use cases to demonstrate value and build alignment. 

Second is building a strong data foundation through investment in governance, cleansing and integration to ensure data quality. 

The third strategy is to implement MLOps to standardise the deployment, monitoring and maintenance of AI models.

Fourth, businesses must upskill their workforce by developing internal AI skills and involving users early in the process. 

Finally, fostering cross-functional collaboration between leadership, IT and operations is essential to align AI initiatives with clear business goals and key performance indicators. 

Success appears to depend less on having the largest infrastructure and more on knowing how to use it effectively.